AlphaMissense Score Interpretation for Missense Variants

What the Score Represents

AlphaMissense is a Google DeepMind model for missense substitutions. It combines protein sequence and structural context to estimate whether an amino acid change is likely to impair protein function.

The model does not score splice variants, truncating variants, in-frame indels, or non-coding variants. Folklore uses BayesDel_noAF and SpliceAI for the formal PP3 and BP4 evidence paths.

How to Read the Score

AlphaMissense scores range from 0 to 1, with higher scores indicating a greater likelihood of pathogenicity. Variants are classified into three categories:

Strengths and Limitations

AlphaMissense's primary strength is that it incorporates protein three-dimensional structural information. While sequence-based tools like SIFT can only assess conservation at a position, AlphaMissense understands whether the amino acid sits in the protein core, on the surface, at an interaction interface, or near a catalytic site. It was trained on human population data and primate conservation data, giving it a human-specific perspective on pathogenicity.

The main limitation is scope: AlphaMissense only predicts impact for missense variants. It does not assess in-frame indels, splice variants, nonsense variants, or any non-coding variants. Its "Ambiguous" category covers a meaningful fraction of all possible missense variants where the model lacks confidence.